summaryrefslogtreecommitdiffstats
path: root/src/video_core/renderer_vulkan
diff options
context:
space:
mode:
Diffstat (limited to 'src/video_core/renderer_vulkan')
-rw-r--r--src/video_core/renderer_vulkan/renderer_vulkan.cpp18
-rw-r--r--src/video_core/renderer_vulkan/vk_blit_screen.cpp6
-rw-r--r--src/video_core/renderer_vulkan/vk_descriptor_pool.cpp7
-rw-r--r--src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp16
-rw-r--r--src/video_core/renderer_vulkan/vk_master_semaphore.h17
-rw-r--r--src/video_core/renderer_vulkan/vk_pipeline_cache.cpp6
-rw-r--r--src/video_core/renderer_vulkan/vk_query_cache.cpp3
-rw-r--r--src/video_core/renderer_vulkan/vk_rasterizer.cpp27
-rw-r--r--src/video_core/renderer_vulkan/vk_scheduler.cpp59
-rw-r--r--src/video_core/renderer_vulkan/vk_scheduler.h13
-rw-r--r--src/video_core/renderer_vulkan/vk_state_tracker.h4
-rw-r--r--src/video_core/renderer_vulkan/vk_swapchain.cpp59
-rw-r--r--src/video_core/renderer_vulkan/vk_swapchain.h21
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.cpp10
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache.h18
-rw-r--r--src/video_core/renderer_vulkan/vk_texture_cache_base.cpp10
16 files changed, 194 insertions, 100 deletions
diff --git a/src/video_core/renderer_vulkan/renderer_vulkan.cpp b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
index 7c9b0d6db..74822814d 100644
--- a/src/video_core/renderer_vulkan/renderer_vulkan.cpp
+++ b/src/video_core/renderer_vulkan/renderer_vulkan.cpp
@@ -97,19 +97,14 @@ RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
Core::Frontend::EmuWindow& emu_window,
Core::Memory::Memory& cpu_memory_, Tegra::GPU& gpu_,
std::unique_ptr<Core::Frontend::GraphicsContext> context_) try
- : RendererBase(emu_window, std::move(context_)),
- telemetry_session(telemetry_session_),
- cpu_memory(cpu_memory_),
- gpu(gpu_),
- library(OpenLibrary()),
+ : RendererBase(emu_window, std::move(context_)), telemetry_session(telemetry_session_),
+ cpu_memory(cpu_memory_), gpu(gpu_), library(OpenLibrary()),
instance(CreateInstance(library, dld, VK_API_VERSION_1_1, render_window.GetWindowInfo().type,
true, Settings::values.renderer_debug.GetValue())),
debug_callback(Settings::values.renderer_debug ? CreateDebugCallback(instance) : nullptr),
surface(CreateSurface(instance, render_window)),
- device(CreateDevice(instance, dld, *surface)),
- memory_allocator(device, false),
- state_tracker(gpu),
- scheduler(device, state_tracker),
+ device(CreateDevice(instance, dld, *surface)), memory_allocator(device, false),
+ state_tracker(gpu), scheduler(device, state_tracker),
swapchain(*surface, device, scheduler, render_window.GetFramebufferLayout().width,
render_window.GetFramebufferLayout().height, false),
blit_screen(cpu_memory, render_window, device, memory_allocator, swapchain, scheduler,
@@ -149,7 +144,7 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
const Layout::FramebufferLayout layout = render_window.GetFramebufferLayout();
swapchain.Create(layout.width, layout.height, is_srgb);
};
- if (swapchain.IsSubOptimal() || swapchain.HasColorSpaceChanged(is_srgb)) {
+ if (swapchain.NeedsRecreation(is_srgb)) {
recreate_swapchain();
}
bool is_outdated;
@@ -164,7 +159,8 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
blit_screen.Recreate();
}
const VkSemaphore render_semaphore = blit_screen.DrawToSwapchain(*framebuffer, use_accelerated);
- scheduler.Flush(render_semaphore);
+ const VkSemaphore present_semaphore = swapchain.CurrentPresentSemaphore();
+ scheduler.Flush(render_semaphore, present_semaphore);
scheduler.WaitWorker();
swapchain.Present(render_semaphore);
diff --git a/src/video_core/renderer_vulkan/vk_blit_screen.cpp b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
index 5c43b8acf..888bc7392 100644
--- a/src/video_core/renderer_vulkan/vk_blit_screen.cpp
+++ b/src/video_core/renderer_vulkan/vk_blit_screen.cpp
@@ -159,11 +159,13 @@ VkSemaphore VKBlitScreen::Draw(const Tegra::FramebufferConfig& framebuffer,
const VAddr framebuffer_addr = framebuffer.address + framebuffer.offset;
const u8* const host_ptr = cpu_memory.GetPointer(framebuffer_addr);
- const size_t size_bytes = GetSizeInBytes(framebuffer);
// TODO(Rodrigo): Read this from HLE
constexpr u32 block_height_log2 = 4;
const u32 bytes_per_pixel = GetBytesPerPixel(framebuffer);
+ const u64 size_bytes{Tegra::Texture::CalculateSize(true, bytes_per_pixel,
+ framebuffer.stride, framebuffer.height,
+ 1, block_height_log2, 0)};
Tegra::Texture::UnswizzleTexture(
mapped_span.subspan(image_offset, size_bytes), std::span(host_ptr, size_bytes),
bytes_per_pixel, framebuffer.width, framebuffer.height, 1, block_height_log2, 0);
@@ -356,7 +358,7 @@ void VKBlitScreen::CreateDescriptorPool() {
void VKBlitScreen::CreateRenderPass() {
const VkAttachmentDescription color_attachment{
.flags = 0,
- .format = swapchain.GetImageFormat(),
+ .format = swapchain.GetImageViewFormat(),
.samples = VK_SAMPLE_COUNT_1_BIT,
.loadOp = VK_ATTACHMENT_LOAD_OP_CLEAR,
.storeOp = VK_ATTACHMENT_STORE_OP_STORE,
diff --git a/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp b/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
index 8e77e4796..d87da2a34 100644
--- a/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
+++ b/src/video_core/renderer_vulkan/vk_descriptor_pool.cpp
@@ -2,6 +2,7 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
+#include <algorithm>
#include <mutex>
#include <span>
#include <vector>
@@ -18,7 +19,6 @@ namespace Vulkan {
// Prefer small grow rates to avoid saturating the descriptor pool with barely used pipelines
constexpr size_t SETS_GROW_RATE = 16;
constexpr s32 SCORE_THRESHOLD = 3;
-constexpr u32 SETS_PER_POOL = 64;
struct DescriptorBank {
DescriptorBankInfo info;
@@ -58,11 +58,12 @@ static DescriptorBankInfo MakeBankInfo(std::span<const Shader::Info> infos) {
static void AllocatePool(const Device& device, DescriptorBank& bank) {
std::array<VkDescriptorPoolSize, 6> pool_sizes;
size_t pool_cursor{};
+ const u32 sets_per_pool = device.GetSetsPerPool();
const auto add = [&](VkDescriptorType type, u32 count) {
if (count > 0) {
pool_sizes[pool_cursor++] = {
.type = type,
- .descriptorCount = count * SETS_PER_POOL,
+ .descriptorCount = count * sets_per_pool,
};
}
};
@@ -77,7 +78,7 @@ static void AllocatePool(const Device& device, DescriptorBank& bank) {
.sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
.pNext = nullptr,
.flags = VK_DESCRIPTOR_POOL_CREATE_FREE_DESCRIPTOR_SET_BIT,
- .maxSets = SETS_PER_POOL,
+ .maxSets = sets_per_pool,
.poolSizeCount = static_cast<u32>(pool_cursor),
.pPoolSizes = std::data(pool_sizes),
}));
diff --git a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
index 7c0f91007..8634c3316 100644
--- a/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
+++ b/src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp
@@ -507,8 +507,9 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
vertex_attributes.push_back({
.location = static_cast<u32>(index),
.binding = 0,
- .format = type == 1 ? VK_FORMAT_R32_SFLOAT
- : type == 2 ? VK_FORMAT_R32_SINT : VK_FORMAT_R32_UINT,
+ .format = type == 1 ? VK_FORMAT_R32_SFLOAT
+ : type == 2 ? VK_FORMAT_R32_SINT
+ : VK_FORMAT_R32_UINT,
.offset = 0,
});
}
@@ -567,12 +568,21 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
if (!vertex_binding_divisors.empty()) {
vertex_input_ci.pNext = &input_divisor_ci;
}
+ const bool has_tess_stages = spv_modules[1] || spv_modules[2];
auto input_assembly_topology = MaxwellToVK::PrimitiveTopology(device, key.state.topology);
if (input_assembly_topology == VK_PRIMITIVE_TOPOLOGY_PATCH_LIST) {
- if (!spv_modules[1] && !spv_modules[2]) {
+ if (!has_tess_stages) {
LOG_WARNING(Render_Vulkan, "Patch topology used without tessellation, using points");
input_assembly_topology = VK_PRIMITIVE_TOPOLOGY_POINT_LIST;
}
+ } else {
+ if (has_tess_stages) {
+ // The Vulkan spec requires patch list IA topology be used with tessellation
+ // shader stages. Forcing it fixes a crash on some drivers
+ LOG_WARNING(Render_Vulkan,
+ "Patch topology not used with tessellation, using patch list");
+ input_assembly_topology = VK_PRIMITIVE_TOPOLOGY_PATCH_LIST;
+ }
}
const VkPipelineInputAssemblyStateCreateInfo input_assembly_ci{
.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
diff --git a/src/video_core/renderer_vulkan/vk_master_semaphore.h b/src/video_core/renderer_vulkan/vk_master_semaphore.h
index 4f8688118..0886b7da8 100644
--- a/src/video_core/renderer_vulkan/vk_master_semaphore.h
+++ b/src/video_core/renderer_vulkan/vk_master_semaphore.h
@@ -21,12 +21,12 @@ public:
/// Returns the current logical tick.
[[nodiscard]] u64 CurrentTick() const noexcept {
- return current_tick.load(std::memory_order_relaxed);
+ return current_tick.load(std::memory_order_acquire);
}
/// Returns the last known GPU tick.
[[nodiscard]] u64 KnownGpuTick() const noexcept {
- return gpu_tick.load(std::memory_order_relaxed);
+ return gpu_tick.load(std::memory_order_acquire);
}
/// Returns the timeline semaphore handle.
@@ -41,12 +41,21 @@ public:
/// Advance to the logical tick and return the old one
[[nodiscard]] u64 NextTick() noexcept {
- return current_tick.fetch_add(1, std::memory_order::relaxed);
+ return current_tick.fetch_add(1, std::memory_order_release);
}
/// Refresh the known GPU tick
void Refresh() {
- gpu_tick.store(semaphore.GetCounter(), std::memory_order_relaxed);
+ u64 this_tick{};
+ u64 counter{};
+ do {
+ this_tick = gpu_tick.load(std::memory_order_acquire);
+ counter = semaphore.GetCounter();
+ if (counter < this_tick) {
+ return;
+ }
+ } while (!gpu_tick.compare_exchange_weak(this_tick, counter, std::memory_order_release,
+ std::memory_order_relaxed));
}
/// Waits for a tick to be hit on the GPU
diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
index a37ca1fdf..eb8b4e08b 100644
--- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp
@@ -281,7 +281,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
.supported_spirv = device.IsKhrSpirv1_4Supported() ? 0x00010400U : 0x00010000U,
.unified_descriptor_binding = true,
.support_descriptor_aliasing = true,
- .support_int8 = true,
+ .support_int8 = device.IsInt8Supported(),
.support_int16 = device.IsShaderInt16Supported(),
.support_int64 = device.IsShaderInt64Supported(),
.support_vertex_instance_id = false,
@@ -325,6 +325,8 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
host_info = Shader::HostTranslateInfo{
.support_float16 = device.IsFloat16Supported(),
.support_int64 = device.IsShaderInt64Supported(),
+ .needs_demote_reorder = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY_KHR ||
+ driver_id == VK_DRIVER_ID_AMD_OPEN_SOURCE_KHR,
};
}
@@ -445,6 +447,8 @@ void PipelineCache::LoadDiskResources(u64 title_id, std::stop_token stop_loading
VideoCommon::LoadPipelines(stop_loading, pipeline_cache_filename, CACHE_VERSION, load_compute,
load_graphics);
+ LOG_INFO(Render_Vulkan, "Total Pipeline Count: {}", state.total);
+
std::unique_lock lock{state.mutex};
callback(VideoCore::LoadCallbackStage::Build, 0, state.total);
state.has_loaded = true;
diff --git a/src/video_core/renderer_vulkan/vk_query_cache.cpp b/src/video_core/renderer_vulkan/vk_query_cache.cpp
index c9cb32d71..259cba156 100644
--- a/src/video_core/renderer_vulkan/vk_query_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_query_cache.cpp
@@ -117,7 +117,8 @@ u64 HostCounter::BlockingQuery() const {
cache.GetScheduler().Wait(tick);
u64 data;
const VkResult query_result = cache.GetDevice().GetLogical().GetQueryResults(
- query.first, query.second, 1, sizeof(data), &data, sizeof(data), VK_QUERY_RESULT_64_BIT);
+ query.first, query.second, 1, sizeof(data), &data, sizeof(data),
+ VK_QUERY_RESULT_64_BIT | VK_QUERY_RESULT_WAIT_BIT);
switch (query_result) {
case VK_SUCCESS:
diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
index 23cef2996..3bcd6d6cc 100644
--- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp
+++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp
@@ -32,7 +32,7 @@
#include "video_core/renderer_vulkan/vk_texture_cache.h"
#include "video_core/renderer_vulkan/vk_update_descriptor.h"
#include "video_core/shader_cache.h"
-#include "video_core/texture_cache/texture_cache.h"
+#include "video_core/texture_cache/texture_cache_base.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -228,9 +228,7 @@ void RasterizerVulkan::Clear() {
};
const u32 color_attachment = regs.clear_buffers.RT;
- const auto attachment_aspect_mask = framebuffer->ImageRanges()[color_attachment].aspectMask;
- const bool is_color_rt = (attachment_aspect_mask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
- if (use_color && is_color_rt) {
+ if (use_color && framebuffer->HasAspectColorBit(color_attachment)) {
VkClearValue clear_value;
std::memcpy(clear_value.color.float32, regs.clear_color, sizeof(regs.clear_color));
@@ -248,12 +246,15 @@ void RasterizerVulkan::Clear() {
return;
}
VkImageAspectFlags aspect_flags = 0;
- if (use_depth) {
+ if (use_depth && framebuffer->HasAspectDepthBit()) {
aspect_flags |= VK_IMAGE_ASPECT_DEPTH_BIT;
}
- if (use_stencil) {
+ if (use_stencil && framebuffer->HasAspectStencilBit()) {
aspect_flags |= VK_IMAGE_ASPECT_STENCIL_BIT;
}
+ if (aspect_flags == 0) {
+ return;
+ }
scheduler.Record([clear_depth = regs.clear_depth, clear_stencil = regs.clear_stencil,
clear_rect, aspect_flags](vk::CommandBuffer cmdbuf) {
VkClearAttachment attachment;
@@ -764,12 +765,7 @@ void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
const Maxwell::StencilOp zpass = regs.stencil_front_op_zpass;
const Maxwell::ComparisonOp compare = regs.stencil_front_func_func;
if (regs.stencil_two_side_enable) {
- scheduler.Record([fail, zfail, zpass, compare](vk::CommandBuffer cmdbuf) {
- cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_AND_BACK, MaxwellToVK::StencilOp(fail),
- MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
- MaxwellToVK::ComparisonOp(compare));
- });
- } else {
+ // Separate stencil op per face
const Maxwell::StencilOp back_fail = regs.stencil_back_op_fail;
const Maxwell::StencilOp back_zfail = regs.stencil_back_op_zfail;
const Maxwell::StencilOp back_zpass = regs.stencil_back_op_zpass;
@@ -784,6 +780,13 @@ void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
MaxwellToVK::StencilOp(back_zfail),
MaxwellToVK::ComparisonOp(back_compare));
});
+ } else {
+ // Front face defines the stencil op of both faces
+ scheduler.Record([fail, zfail, zpass, compare](vk::CommandBuffer cmdbuf) {
+ cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_AND_BACK, MaxwellToVK::StencilOp(fail),
+ MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
+ MaxwellToVK::ComparisonOp(compare));
+ });
}
}
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.cpp b/src/video_core/renderer_vulkan/vk_scheduler.cpp
index 4840962de..0c11c814f 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.cpp
+++ b/src/video_core/renderer_vulkan/vk_scheduler.cpp
@@ -43,26 +43,19 @@ VKScheduler::VKScheduler(const Device& device_, StateTracker& state_tracker_)
command_pool{std::make_unique<CommandPool>(*master_semaphore, device)} {
AcquireNewChunk();
AllocateWorkerCommandBuffer();
- worker_thread = std::thread(&VKScheduler::WorkerThread, this);
+ worker_thread = std::jthread([this](std::stop_token token) { WorkerThread(token); });
}
-VKScheduler::~VKScheduler() {
- {
- std::lock_guard lock{work_mutex};
- quit = true;
- }
- work_cv.notify_all();
- worker_thread.join();
-}
+VKScheduler::~VKScheduler() = default;
-void VKScheduler::Flush(VkSemaphore semaphore) {
- SubmitExecution(semaphore);
+void VKScheduler::Flush(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
+ SubmitExecution(signal_semaphore, wait_semaphore);
AllocateNewContext();
}
-void VKScheduler::Finish(VkSemaphore semaphore) {
+void VKScheduler::Finish(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
const u64 presubmit_tick = CurrentTick();
- SubmitExecution(semaphore);
+ SubmitExecution(signal_semaphore, wait_semaphore);
WaitWorker();
Wait(presubmit_tick);
AllocateNewContext();
@@ -135,7 +128,7 @@ bool VKScheduler::UpdateGraphicsPipeline(GraphicsPipeline* pipeline) {
return true;
}
-void VKScheduler::WorkerThread() {
+void VKScheduler::WorkerThread(std::stop_token stop_token) {
Common::SetCurrentThreadName("yuzu:VulkanWorker");
do {
if (work_queue.empty()) {
@@ -144,8 +137,8 @@ void VKScheduler::WorkerThread() {
std::unique_ptr<CommandChunk> work;
{
std::unique_lock lock{work_mutex};
- work_cv.wait(lock, [this] { return !work_queue.empty() || quit; });
- if (quit) {
+ work_cv.wait(lock, stop_token, [this] { return !work_queue.empty(); });
+ if (stop_token.stop_requested()) {
continue;
}
work = std::move(work_queue.front());
@@ -158,7 +151,7 @@ void VKScheduler::WorkerThread() {
}
std::lock_guard reserve_lock{reserve_mutex};
chunk_reserve.push_back(std::move(work));
- } while (!quit);
+ } while (!stop_token.stop_requested());
}
void VKScheduler::AllocateWorkerCommandBuffer() {
@@ -171,37 +164,41 @@ void VKScheduler::AllocateWorkerCommandBuffer() {
});
}
-void VKScheduler::SubmitExecution(VkSemaphore semaphore) {
+void VKScheduler::SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore) {
EndPendingOperations();
InvalidateState();
const u64 signal_value = master_semaphore->NextTick();
- Record([semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
+ Record([signal_semaphore, wait_semaphore, signal_value, this](vk::CommandBuffer cmdbuf) {
cmdbuf.End();
-
- const u32 num_signal_semaphores = semaphore ? 2U : 1U;
-
- const u64 wait_value = signal_value - 1;
- const VkPipelineStageFlags wait_stage_mask = VK_PIPELINE_STAGE_ALL_COMMANDS_BIT;
-
const VkSemaphore timeline_semaphore = master_semaphore->Handle();
+
+ const u32 num_signal_semaphores = signal_semaphore ? 2U : 1U;
const std::array signal_values{signal_value, u64(0)};
- const std::array signal_semaphores{timeline_semaphore, semaphore};
+ const std::array signal_semaphores{timeline_semaphore, signal_semaphore};
+
+ const u32 num_wait_semaphores = wait_semaphore ? 2U : 1U;
+ const std::array wait_values{signal_value - 1, u64(1)};
+ const std::array wait_semaphores{timeline_semaphore, wait_semaphore};
+ static constexpr std::array<VkPipelineStageFlags, 2> wait_stage_masks{
+ VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+ VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
+ };
const VkTimelineSemaphoreSubmitInfoKHR timeline_si{
.sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO_KHR,
.pNext = nullptr,
- .waitSemaphoreValueCount = 1,
- .pWaitSemaphoreValues = &wait_value,
+ .waitSemaphoreValueCount = num_wait_semaphores,
+ .pWaitSemaphoreValues = wait_values.data(),
.signalSemaphoreValueCount = num_signal_semaphores,
.pSignalSemaphoreValues = signal_values.data(),
};
const VkSubmitInfo submit_info{
.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO,
.pNext = &timeline_si,
- .waitSemaphoreCount = 1,
- .pWaitSemaphores = &timeline_semaphore,
- .pWaitDstStageMask = &wait_stage_mask,
+ .waitSemaphoreCount = num_wait_semaphores,
+ .pWaitSemaphores = wait_semaphores.data(),
+ .pWaitDstStageMask = wait_stage_masks.data(),
.commandBufferCount = 1,
.pCommandBuffers = cmdbuf.address(),
.signalSemaphoreCount = num_signal_semaphores,
diff --git a/src/video_core/renderer_vulkan/vk_scheduler.h b/src/video_core/renderer_vulkan/vk_scheduler.h
index cf39a2363..85fc1712f 100644
--- a/src/video_core/renderer_vulkan/vk_scheduler.h
+++ b/src/video_core/renderer_vulkan/vk_scheduler.h
@@ -34,10 +34,10 @@ public:
~VKScheduler();
/// Sends the current execution context to the GPU.
- void Flush(VkSemaphore semaphore = nullptr);
+ void Flush(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
/// Sends the current execution context to the GPU and waits for it to complete.
- void Finish(VkSemaphore semaphore = nullptr);
+ void Finish(VkSemaphore signal_semaphore = nullptr, VkSemaphore wait_semaphore = nullptr);
/// Waits for the worker thread to finish executing everything. After this function returns it's
/// safe to touch worker resources.
@@ -187,11 +187,11 @@ private:
GraphicsPipeline* graphics_pipeline = nullptr;
};
- void WorkerThread();
+ void WorkerThread(std::stop_token stop_token);
void AllocateWorkerCommandBuffer();
- void SubmitExecution(VkSemaphore semaphore);
+ void SubmitExecution(VkSemaphore signal_semaphore, VkSemaphore wait_semaphore);
void AllocateNewContext();
@@ -212,7 +212,6 @@ private:
vk::CommandBuffer current_cmdbuf;
std::unique_ptr<CommandChunk> chunk;
- std::thread worker_thread;
State state;
@@ -224,9 +223,9 @@ private:
std::vector<std::unique_ptr<CommandChunk>> chunk_reserve;
std::mutex reserve_mutex;
std::mutex work_mutex;
- std::condition_variable work_cv;
+ std::condition_variable_any work_cv;
std::condition_variable wait_cv;
- std::atomic_bool quit{};
+ std::jthread worker_thread;
};
} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_state_tracker.h b/src/video_core/renderer_vulkan/vk_state_tracker.h
index 5f78f6950..d90935f52 100644
--- a/src/video_core/renderer_vulkan/vk_state_tracker.h
+++ b/src/video_core/renderer_vulkan/vk_state_tracker.h
@@ -110,10 +110,6 @@ public:
return Exchange(Dirty::DepthTestEnable, false);
}
- bool TouchDepthBoundsEnable() {
- return Exchange(Dirty::DepthBoundsEnable, false);
- }
-
bool TouchDepthWriteEnable() {
return Exchange(Dirty::DepthWriteEnable, false);
}
diff --git a/src/video_core/renderer_vulkan/vk_swapchain.cpp b/src/video_core/renderer_vulkan/vk_swapchain.cpp
index d990eefba..8972a6921 100644
--- a/src/video_core/renderer_vulkan/vk_swapchain.cpp
+++ b/src/video_core/renderer_vulkan/vk_swapchain.cpp
@@ -9,6 +9,7 @@
#include "common/assert.h"
#include "common/logging/log.h"
+#include "common/settings.h"
#include "core/core.h"
#include "core/frontend/framebuffer_layout.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
@@ -20,16 +21,15 @@ namespace Vulkan {
namespace {
-VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats, bool srgb) {
+VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats) {
if (formats.size() == 1 && formats[0].format == VK_FORMAT_UNDEFINED) {
VkSurfaceFormatKHR format;
format.format = VK_FORMAT_B8G8R8A8_UNORM;
format.colorSpace = VK_COLOR_SPACE_SRGB_NONLINEAR_KHR;
return format;
}
- const auto& found = std::find_if(formats.begin(), formats.end(), [srgb](const auto& format) {
- const auto request_format = srgb ? VK_FORMAT_B8G8R8A8_SRGB : VK_FORMAT_B8G8R8A8_UNORM;
- return format.format == request_format &&
+ const auto& found = std::find_if(formats.begin(), formats.end(), [](const auto& format) {
+ return format.format == VK_FORMAT_B8G8R8A8_UNORM &&
format.colorSpace == VK_COLOR_SPACE_SRGB_NONLINEAR_KHR;
});
return found != formats.end() ? *found : formats[0];
@@ -37,8 +37,19 @@ VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats,
VkPresentModeKHR ChooseSwapPresentMode(vk::Span<VkPresentModeKHR> modes) {
// Mailbox doesn't lock the application like fifo (vsync), prefer it
- const auto found = std::find(modes.begin(), modes.end(), VK_PRESENT_MODE_MAILBOX_KHR);
- return found != modes.end() ? *found : VK_PRESENT_MODE_FIFO_KHR;
+ const auto found_mailbox = std::find(modes.begin(), modes.end(), VK_PRESENT_MODE_MAILBOX_KHR);
+ if (found_mailbox != modes.end()) {
+ return VK_PRESENT_MODE_MAILBOX_KHR;
+ }
+ if (Settings::values.disable_fps_limit.GetValue()) {
+ // FIFO present mode locks the framerate to the monitor's refresh rate,
+ // Find an alternative to surpass this limitation if FPS is unlocked.
+ const auto found_imm = std::find(modes.begin(), modes.end(), VK_PRESENT_MODE_IMMEDIATE_KHR);
+ if (found_imm != modes.end()) {
+ return VK_PRESENT_MODE_IMMEDIATE_KHR;
+ }
+ }
+ return VK_PRESENT_MODE_FIFO_KHR;
}
VkExtent2D ChooseSwapExtent(const VkSurfaceCapabilitiesKHR& capabilities, u32 width, u32 height) {
@@ -107,14 +118,12 @@ void VKSwapchain::AcquireNextImage() {
}
void VKSwapchain::Present(VkSemaphore render_semaphore) {
- const VkSemaphore present_semaphore{*present_semaphores[frame_index]};
- const std::array<VkSemaphore, 2> semaphores{present_semaphore, render_semaphore};
const auto present_queue{device.GetPresentQueue()};
const VkPresentInfoKHR present_info{
.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
.pNext = nullptr,
- .waitSemaphoreCount = render_semaphore ? 2U : 1U,
- .pWaitSemaphores = semaphores.data(),
+ .waitSemaphoreCount = render_semaphore ? 1U : 0U,
+ .pWaitSemaphores = &render_semaphore,
.swapchainCount = 1,
.pSwapchains = swapchain.address(),
.pImageIndices = &image_index,
@@ -145,8 +154,8 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
const auto formats{physical_device.GetSurfaceFormatsKHR(surface)};
const auto present_modes{physical_device.GetSurfacePresentModesKHR(surface)};
- const VkSurfaceFormatKHR surface_format{ChooseSwapSurfaceFormat(formats, srgb)};
- const VkPresentModeKHR present_mode{ChooseSwapPresentMode(present_modes)};
+ const VkSurfaceFormatKHR surface_format{ChooseSwapSurfaceFormat(formats)};
+ present_mode = ChooseSwapPresentMode(present_modes);
u32 requested_image_count{capabilities.minImageCount + 1};
if (capabilities.maxImageCount > 0 && requested_image_count > capabilities.maxImageCount) {
@@ -180,6 +189,17 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
swapchain_ci.queueFamilyIndexCount = static_cast<u32>(queue_indices.size());
swapchain_ci.pQueueFamilyIndices = queue_indices.data();
}
+ static constexpr std::array view_formats{VK_FORMAT_B8G8R8A8_UNORM, VK_FORMAT_B8G8R8A8_SRGB};
+ VkImageFormatListCreateInfo format_list{
+ .sType = VK_STRUCTURE_TYPE_IMAGE_FORMAT_LIST_CREATE_INFO_KHR,
+ .pNext = nullptr,
+ .viewFormatCount = static_cast<u32>(view_formats.size()),
+ .pViewFormats = view_formats.data(),
+ };
+ if (device.IsKhrSwapchainMutableFormatEnabled()) {
+ format_list.pNext = std::exchange(swapchain_ci.pNext, &format_list);
+ swapchain_ci.flags |= VK_SWAPCHAIN_CREATE_MUTABLE_FORMAT_BIT_KHR;
+ }
// Request the size again to reduce the possibility of a TOCTOU race condition.
const auto updated_capabilities = physical_device.GetSurfaceCapabilitiesKHR(surface);
swapchain_ci.imageExtent = ChooseSwapExtent(updated_capabilities, width, height);
@@ -188,10 +208,11 @@ void VKSwapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities,
extent = swapchain_ci.imageExtent;
current_srgb = srgb;
+ current_fps_unlocked = Settings::values.disable_fps_limit.GetValue();
images = swapchain.GetImages();
image_count = static_cast<u32>(images.size());
- image_format = surface_format.format;
+ image_view_format = srgb ? VK_FORMAT_B8G8R8A8_SRGB : VK_FORMAT_B8G8R8A8_UNORM;
}
void VKSwapchain::CreateSemaphores() {
@@ -207,7 +228,7 @@ void VKSwapchain::CreateImageViews() {
.flags = 0,
.image = {},
.viewType = VK_IMAGE_VIEW_TYPE_2D,
- .format = image_format,
+ .format = image_view_format,
.components =
{
.r = VK_COMPONENT_SWIZZLE_IDENTITY,
@@ -240,4 +261,14 @@ void VKSwapchain::Destroy() {
swapchain.reset();
}
+bool VKSwapchain::HasFpsUnlockChanged() const {
+ return current_fps_unlocked != Settings::values.disable_fps_limit.GetValue();
+}
+
+bool VKSwapchain::NeedsPresentModeUpdate() const {
+ // Mailbox present mode is the ideal for all scenarios. If it is not available,
+ // A different present mode is needed to support unlocked FPS above the monitor's refresh rate.
+ return present_mode != VK_PRESENT_MODE_MAILBOX_KHR && HasFpsUnlockChanged();
+}
+
} // namespace Vulkan
diff --git a/src/video_core/renderer_vulkan/vk_swapchain.h b/src/video_core/renderer_vulkan/vk_swapchain.h
index 35c2cdc14..61a6d959e 100644
--- a/src/video_core/renderer_vulkan/vk_swapchain.h
+++ b/src/video_core/renderer_vulkan/vk_swapchain.h
@@ -33,6 +33,11 @@ public:
/// Presents the rendered image to the swapchain.
void Present(VkSemaphore render_semaphore);
+ /// Returns true when the swapchain needs to be recreated.
+ bool NeedsRecreation(bool is_srgb) const {
+ return HasColorSpaceChanged(is_srgb) || IsSubOptimal() || NeedsPresentModeUpdate();
+ }
+
/// Returns true when the color space has changed.
bool HasColorSpaceChanged(bool is_srgb) const {
return current_srgb != is_srgb;
@@ -68,8 +73,12 @@ public:
return *image_views[index];
}
- VkFormat GetImageFormat() const {
- return image_format;
+ VkFormat GetImageViewFormat() const {
+ return image_view_format;
+ }
+
+ VkSemaphore CurrentPresentSemaphore() const {
+ return *present_semaphores[frame_index];
}
private:
@@ -80,6 +89,10 @@ private:
void Destroy();
+ bool HasFpsUnlockChanged() const;
+
+ bool NeedsPresentModeUpdate() const;
+
const VkSurfaceKHR surface;
const Device& device;
VKScheduler& scheduler;
@@ -96,10 +109,12 @@ private:
u32 image_index{};
u32 frame_index{};
- VkFormat image_format{};
+ VkFormat image_view_format{};
VkExtent2D extent{};
+ VkPresentModeKHR present_mode{};
bool current_srgb{};
+ bool current_fps_unlocked{};
bool is_outdated{};
bool is_suboptimal{};
};
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.cpp b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
index 8e029bcb3..06c5fb867 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.cpp
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.cpp
@@ -19,6 +19,9 @@
#include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
#include "video_core/renderer_vulkan/vk_texture_cache.h"
+#include "video_core/texture_cache/formatter.h"
+#include "video_core/texture_cache/samples_helper.h"
+#include "video_core/texture_cache/util.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -125,7 +128,7 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
const auto format_info = MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, false, format);
VkImageCreateFlags flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
if (info.type == ImageType::e2D && info.resources.layers >= 6 &&
- info.size.width == info.size.height) {
+ info.size.width == info.size.height && !device.HasBrokenCubeImageCompability()) {
flags |= VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT;
}
if (info.type == ImageType::e3D) {
@@ -1184,9 +1187,12 @@ Framebuffer::Framebuffer(TextureCacheRuntime& runtime, std::span<ImageView*, NUM
renderpass_key.depth_format = depth_buffer->format;
num_layers = std::max(num_layers, depth_buffer->range.extent.layers);
images[num_images] = depth_buffer->ImageHandle();
- image_ranges[num_images] = MakeSubresourceRange(depth_buffer);
+ const VkImageSubresourceRange subresource_range = MakeSubresourceRange(depth_buffer);
+ image_ranges[num_images] = subresource_range;
samples = depth_buffer->Samples();
++num_images;
+ has_depth = (subresource_range.aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0;
+ has_stencil = (subresource_range.aspectMask & VK_IMAGE_ASPECT_STENCIL_BIT) != 0;
} else {
renderpass_key.depth_format = PixelFormat::Invalid;
}
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h
index 0b73d55f8..b09c468e4 100644
--- a/src/video_core/renderer_vulkan/vk_texture_cache.h
+++ b/src/video_core/renderer_vulkan/vk_texture_cache.h
@@ -4,12 +4,12 @@
#pragma once
-#include <compare>
#include <span>
#include "shader_recompiler/shader_info.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
-#include "video_core/texture_cache/texture_cache.h"
+#include "video_core/texture_cache/image_view_base.h"
+#include "video_core/texture_cache/texture_cache_base.h"
#include "video_core/vulkan_common/vulkan_memory_allocator.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
@@ -232,6 +232,18 @@ public:
return image_ranges;
}
+ [[nodiscard]] bool HasAspectColorBit(size_t index) const noexcept {
+ return (image_ranges.at(index).aspectMask & VK_IMAGE_ASPECT_COLOR_BIT) != 0;
+ }
+
+ [[nodiscard]] bool HasAspectDepthBit() const noexcept {
+ return has_depth;
+ }
+
+ [[nodiscard]] bool HasAspectStencilBit() const noexcept {
+ return has_stencil;
+ }
+
private:
vk::Framebuffer framebuffer;
VkRenderPass renderpass{};
@@ -241,6 +253,8 @@ private:
u32 num_images = 0;
std::array<VkImage, 9> images{};
std::array<VkImageSubresourceRange, 9> image_ranges{};
+ bool has_depth{};
+ bool has_stencil{};
};
struct TextureCacheParams {
diff --git a/src/video_core/renderer_vulkan/vk_texture_cache_base.cpp b/src/video_core/renderer_vulkan/vk_texture_cache_base.cpp
new file mode 100644
index 000000000..44e688342
--- /dev/null
+++ b/src/video_core/renderer_vulkan/vk_texture_cache_base.cpp
@@ -0,0 +1,10 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/renderer_vulkan/vk_texture_cache.h"
+#include "video_core/texture_cache/texture_cache.h"
+
+namespace VideoCommon {
+template class VideoCommon::TextureCache<Vulkan::TextureCacheParams>;
+}
generated by cgit v1.2.3 (git 2.25.1) at 2024-07-05 12:20:21 +0200